Didn't AMD patent this a few months ago?
Are they dumping 2 vector ALUs and replacing them with 2 scalar ALUs because the vector ALUs weren't utilized in a 16 wide SIMD design?
Is everyone dead or busy shitposting in handheld threads?
>>56953127
It seems pretty clear that this is just a way for shaders to operate on any vector width from 16 down to 1 with no wastage.
The bane of SIMT is repacking warps/wavefronts as thread flows diverge from data-dependent branching, and I guess this is a way to make repacking easier or at least being able to gate off unused ALU slots to save power.
>>56953075
This has been floating around for a while, but who the fuck knows whether the curryniggers actually have anything real here.
I'm still waiting on hard confirmation that Vega will do double-rate fp16.
If anything this is the single largest architecture change for AMD's GPU since TeraScale > GCN
So it's either gonna be fantastic or plain shit, either way it's gonna be interesting.
>>56953406
It's a neat concept, but who the fuck even knows if this will actually be seen in Vega or even Navi honestly.
GPU things that have already let me down this year:
> HBM2 in Spring 2016!
> Polaris 10 will be 120 W!
> 1070 will beat a 980 Ti and cost $350 like the 970!
> Vega will have 1TB/s bandwidth and a fuckton of shaders!
> New Titan/1080 Ti will be GP100 based or at least be ~600mm^2 monsters!
> DX12/Vulkan titles will be everywhere!
Maybe Vega will be great, but it'll have to be a miracle to make up for letting Nvidia go turbo-kike in this year.
>>56953552
If they're planning on taking on the 1080ti with the new Vega they'll need major change if they want to keep 4096 shaders.
So yeah, I believe this vector+scalar design is gonna be used.
>>56953552
I picked up a G1 470 4GB for $200 when the RX480s were ridiculously priced. Had had no graphics card for a couple months, couldn't be fucked waiting.
That card will last until the 7nm GPUs arrive in a couple years. GCN4 is a good arch considering it's used in the PS4 and XB1 with a few extra features like texture compression and tessellation improvement.
At 7nm you'll get all the cool new stuff like HBM for cheap
>>56953838
AMD clearly needs to fix something if 4 TFLOPS 1060s compete at the same level as if not higher than 6 TFLOPS 480s.
Otherwise a 12 TFLOPS GCN card wouldn't fare well against a 10-11 TFLOPS Pascal card.
>>56954015
AMD TFLOP ails come from ALU underutilazation and high overheard drivers in DX11, it has no problems with either in DX12/Vulkan and it shows.
Hopefully at least one of them will be fixed with the new arch
>>56954064
GCN suffers death by a thousand cuts
> atrocious DX11 overhead and threading design
> bottlenecked by geometry/tessellation/ROPs in most games thanks to TWIMTBP bullshit
> dubious efficiency of the caching and shader allocation subsystems
> unified core clock for poor overclocking/clock scaling
> less efficient memory compression than Maxwell/Pascal
> whole frame instead of tiled delayed rasterization has inferior cache locality characteristics
> ...
Some of the few new Vulkan/DX12 titles have done well for AMD, but the only thing that Polaris really tried to fix over 28nm-gen chips was adding shader instruction prefetch and adding a respectable amount of L2 cache.
There's a LOT of things that AMD can be addressing in Vega/Navi.
>>56954444
I thought everything after Hawaii had tile based rendering.
>>56954922
unless you have source saying so, I've only heard otherwise.
in the end it may not matter much depending on renderer trends. for example, Doom 4 does its lighting and texturing passes on screen-space volumetric bins (something like 8x4x24), which enforces locality at the software level in an intelligent manner.